system.yaml

model: depth-anything/DA3-BASE
device: cuda
process_res: 504
torch_threads: 2
max_tracking_gap_s: 2.0
shutdown_drain_s: 3.0
selected_queue_timeout_s: 2.0
diagnostics_period_s: 1.0
openvins:
  configs:
    euroc: config/euroc/estimator_config.yaml
    tumvi: config/tumvi/estimator_config.yaml
    ori: config/ori/estimator_config.yaml
    realsense: config/realsense/estimator_config.yaml
    vcu_rvi: config/vcu_rvi/estimator_config.yaml
    phone: config/phone/estimator_config.yaml
    custom: config/phone/estimator_config.yaml
  overrides:
    verbosity: WARNING
    calib_cam_extrinsics: true
    calib_cam_intrinsics: false
    calib_cam_timeoffset: false
    init_window_time: 2.0
    init_dyn_use: true
    init_dyn_mle_opt_calib: true
    init_dyn_mle_max_threads: 2
    init_dyn_mle_max_iter: 20
    init_dyn_mle_max_time: 0.05
    init_dyn_num_pose: 5
    init_dyn_min_deg: 15.0
    init_dyn_bias_g: [0.0, 0.0, 0.0]
    init_dyn_bias_a: [0.0, 0.0, 0.0]
calibration:
  mode: supplied
  lever_arm_prior_m: 0.05     # sigma of the p_CinI prior in the feed-forward initializer
  rectifier_refresh_px: 2.0   # free mode: rebuild the DA3 rectifier when the filter's K/D move a corner by more
  priors:                     # initial calibration sigmas written into the assisted filter config (free modes)
    rotation_deg: 15.0     # hand-eye start measured 6-11 deg on M1 seed 0 (v1 shadow study: 3-12 deg); 3 deg diverged on MH_01
    translation_m: 0.10
    focal_px: 20.0
    distortion: 0.1
    time_offset_s: 0.02
assistance:
  enabled: true
  solver: joint
  validate: true
  native_priority: true
  budget_s: 30.0           # free modes need several windows of rotation before a hand-eye release
  window_frames: 5
  sample_period_s: 0.2 
  feedforward: true
  legacy_rotation_proposal: false
  ff:                       # init/feedforward.py default_config keys
    points_per_frame: 100
    conf_quantile: 0.25
    ransac_iterations: 100
    ransac_threshold: 0.01
    min_inlier_fraction: 0.4
    min_frames_covered: 3
    max_median_error: 0.05
    estimate_lever_arm: false   # held at the prior (zero): measured 0.2-3.9 m when free; the filter calibrates it
  attempt_period_s: 0.2
  max_windows: 10          # the joint solve re-preintegrates every window per LM trial; 10 keeps a release under ~1 s
  imu_history_s: 10.0
  history_frames: 160      # 8 s at 20 Hz: must cover the oldest admissible candidate (max_candidate_age_s) plus its window
  shadow_queue_frames: 20
  candidate_preroll_s: null
  shadow_continuous: false
  shadow_period_s: 2.0
  max_candidate_age_s: 5.0
  candidate_timeout_s: 3.0
  max_attempts: 5
  max_catchup_lag_s: 0.15
  warm_frames: 5
  max_bias_norm_rad_s: 0.3
  joint_max_iterations: 40  # LM iterations of the joint hand-eye/bias solve (v1: 15)
  pair_angle_tol_rad_free: 0.1   # hand-eye pair angle-agreement gate in free mode (rectified default 0.05)
  validation_rms_deg_free: 3.0
  validation_max_deg_free: 8.0
  validation_rms_deg: 1.0
  validation_max_deg: 3.0
  gates:
    joint_min_axis_coverage: 0.02
    joint_min_sigma_x_given_b: 1.0e-6
    joint_min_sigma_bb: 1.0e-6
    joint_min_pairs: 3
mapping:
  enabled: true
  window_frames: 5
  max_task_age_s: 2.0
  keyframe_period_s: 0.5
  max_active_submaps: 24        # resident DENSE geometry only; pose nodes are kept
  max_graph_nodes: 5000         # hard bound on the back-end; older nodes freeze
  min_baseline_m: 0.05
  max_alignment_rmse_m: 0.15
  voxel_m: 0.04
  pixel_step: 4
  da3_pose_conditioning: true
  da3_conditioned_max_log_scale: 0.3
  frame_scale_refinement: true
  frame_scale_absolute: false
  frame_scale_abs_prior_sigma: 0.1
  frame_scale_pairs: 2            # match frame i with i+1 .. i+pairs
  frame_scale_min_matches: 30
  frame_scale_prior_sigma: 1.0    # weak prior s_i = 1 so an unmatched frame is left alone
  frame_scale_cut: 3.0            # rows beyond this many robust sigmas are dropped
  frame_scale_rounds: 8
  shared_frame_scale: true
  depth_filter: true
  depth_min_m: 0.15
  depth_max_m: 20.0
  depth_conf_quantile: 0.2
  depth_edge_relative: 0.15
  depth_consistency_relative: 0.10
  depth_min_support: 1
  export_every_submaps: 1
  correction_nodes: 1
  correction_jump_m: 0.05
  correction_jump_deg: 2.0
  correction_ramp_s: 1.0
  graph_iterations: 8          # only runs when a loop disagrees with the odometry chain
  loops_enabled: true
  feature_backend: xfeat_lighterglue
  xfeat_mnn_min_cossim: 0.82     # xfeat_mnn backend and appearance retrieval threshold
  xfeat_path: thirdparty/accelerated_features
  xfeat_features: 2048
  sparse_alignment: true
  sparse_adjacent: true    # adjacent-submap point factors; loops-only ablation turns this off
  adjacent_factor: point   # point (N 3D residuals) or pose (rigid PnP control)
  sparse_max_points: 80
  sparse_sigma_m: 0.10            # isotropic floor on a matched 3D correspondence
  sparse_sigma_radial: 0.10
  sparse_sigma_lateral: 0.01
  overlap_frames: 2
  orb_features: 1000
  max_loop_candidates: 4
  # spatial | appearance | union; which retrieval proposed a loop is recorded per loop.
  loop_retrieval: spatial
  loop_retrieval_budget: 512
  loop_retrieval_min_mnn: 25
  loop_min_separation_s: 5.0
  loop_search_radius_m: 8.0     # spatial gate, not a recency window; states the search scope
  loop_min_matches: 50
  loop_min_inliers: 30
  loop_min_inlier_ratio: 0.25
  loop_reprojection_px: 2.0
  loop_depth_relative_error: 0.2
  loop_max_rotation_deg: 30.0    # absolute ceiling; the normalized gate below usually binds
  loop_max_translation_m: 3.0
  loop_max_innovation_sigma: 3.0
  loop_min_drift_ratio: 1.0
  # False-loop robustness probe (E7). 0 disables; every N submaps one coherent wrong loop.
  inject_false_loop_every: 0
  inject_false_loop_rotation_deg: 10.0
  inject_false_loop_translation_m: 1.0
  inject_seed: 0
  odometry_sigma_floor_m: 0.02
  odometry_sigma_floor_deg: 0.5
  odometry_drift_m_per_m: 0.005
  odometry_drift_deg_per_m: 0.5
  loop_sigma_m: 0.25
  loop_sigma_deg: 5.0
  loop_scale_relative_error: 0.10
  scale_coupling: true
  graph_nodes: gravity
  scale_anchors: true             # per-node log-scale prior to the window fit (E4 arm 7 drops it)
  overlap_scale_factors: true     # relative depth-scale observations from overlaps (E4 arm 8 drops them)
  # C3 selective geometry-to-pose admission (negative result; E4 control only).
  selective_admission: false
  admission_min_information: 1.0
  admission_max_condition: 1.0e4
  admission_heldout_max: 2.0
  admission_drift_sigma: 3.0
  admission_confirm_m: 0.05
  admission_confirm_deg: 2.0
  scale_prior_sigma_log: 0.10      # trust in one window's own baseline fit
  scale_sigma_log: 0.03            # trust in an overlap relative-scale measurement
  scale_max_relative_error: 0.5    # widest depth disagreement still usable as a measurement
  scale_overlap_min_points: 200    # fewest overlapping samples that make a measurement